A 3DoF Pose Estimation Method for Multi-Trolley from a Single RGB Image

3DoF pose estimation for multi-trolley is pivotal in Automatic Collection Trolley Robot System. Because of the challenge of occlusion and low-resolution problem, the existing methods do not solve these problems well. In this paper we propose a 3DoF pose estimation method for multi-trolley based on the keypoints from a single RGB image. The proposed method is divided into three steps: Firstly, YOLOv3 is used to detect the trolley in a RGB image and crop the trolley individually. Secondly a new network including two parts is proposed for detecting the 2D keypoints of the trolley. One part called DetectNet is used for detecting easy keypoints, while the other part called HardNet is for detecting hard keypoints under occlusion followed by post-processing for finetuning. Finally, 3DoF pose estimation of a trolley can be obtained by the corresponding relationship between the 2D keypoints and the real 3D keypoints in trolley 3D model. Experiments have been carried out to validate the proposed model. The results show that our approach is accuracy and robust for 3DoF pose estimation of trolley.